Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
AI+EMFarron Wallace NOAA Fisheries SEFSC
SEFSC Innovation ProjectIn Collaboration
AFSC, University of Washington, NGO’s and the Fishing Industry
AI?
Proof of Concept
• 4 types of rockfish
Blackspotted Rougheye
Shortraker Thornyhead
Test on Rockfish Dataset
• Training (used previous ~15,000 training dataset)• Blackspotted (59), Rougheye (15), Shortraker (21), Thornyhead (9).
• Testing• Blackspotted (20), Rougheye (5), Shortraker (7), Thornyhead (4).
• Confusion Matrix (Accuracy=91.7%)
Blackspotted Rougheye Shortraker Thornyhead
Blackspotted 18 1 0 0
Rougheye 0 4 0 0
Shortraker 2 0 7 0
Thornyhead 0 0 0 4
Machine Vision Systems (IMS)ØRemote monitoring systems that support the
latest developments in machine learning (AI) to improve system functionality, reliability and timeliness of data.
ØAn automated system that can see, learn, process and transmit data
Cost Categories for Developing Machine Vision?
1. Collect Imagery (EM, Chute, Belt Systems, cameras)
2. Annotate imagery to create a rich training dataset.
3. Feed these data in ML algorithm(s).
4. Hardware development and implementation.
Iterative Development CycleInnovateSystem
Hardware
Deploy into Fishery or
Survey
CollectTraining Datasets
Feed intoMachine Learning
ProgressAutomation
AgileProductManagement
Methods based on iterative and incremental development
and testing
ValidationDatasets
QA/QC
Proof ofConcept
Rich and complexdataset
Guidelines for Developing Automation1. The scale of the data needs to reasonably capture the complexity of the problem.
2. Big data must be accurate.
3. At minimum you need an ML algorithm that captures the discriminating information.
Guidelines for Developing Automation4. The more complex the algorithm the greater the processing power/time required.
5. There is a delicate balance in dedicating resources between Algorithm development and collecting more data.
6. Use multiple sources of information (human versus machine) to validate your system
UW Annotation Software
How many annotations do I need for training?
ML
Algo
rithm
Perf
orm
ance
Number of Annotations (1,000)
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 50 100 150 200 250 300 350 400 450 500
Learning Curve
COMPLEXITY
Large Annotated Datasets Required: Longline Fisheries
Ø35,300,000 Ø50,000 Ø706 Ø14
How do you annotate all those images? Semi-automated annotation tools Strategic sub-sampling of dataQuery Learning/Unsupervised Learning (DL)
# frames collected from 184 hauls# frames processed per week# of weeks to process# of years to process
•
Metric Learning with Multiple Representative Features
• To deal with intra-class variance• For class !, define " representative feature vectors #$,& for intra-class variance:'( )* = max min& 1 )* − #$,& − min
3,&41 )* − #3,&4 + 6, 0
#$,&
#3,&4
)*
)*89
)*:9
class ! class ;
Identifying Catch Events
Red: shipGreen: sea
3D bounding box for size estimation
15
What’s Next?• Distribute the knowledge and design of systems that
collect and analyze imagery.
• Provide access to open source Machine Learning algorithms to promote development of automation in other fisheries (VIAMI, GitHub)
• Provide Access to machine vision Stereo Camera systems to promote automation and development of 3D vision systems to estimate size and volume measurements.
Questions?
17
The NOAA AI Strategy will dramatically expand the application of artificial intelligence (AI) in every NOAA mission area by improving the efficiency, effectiveness and coordination of AI development and usage across the agency.
AI methods will provide transformative advancements in the quality and timeliness of NOAA science, products, and services: are already demonstrating significant improvements in performance and skill at vastly reduced costs and compute time.